Requstory Lab · free tools and open researchSuppliers and buyers: the main site
requstory

The AI support agent that never handed back to a person

A small online shop gave first-line support to an AI agent. It worked well until a customer asked for something the agent could not do.

Illustrative story: a composite of how teams use Requstory. No real company or person is named or quoted, and outcomes are what the product shows, not measured business results.

Before: the AI agent's refund step ends with nobody picking it up. After: the agent hands refunds to a person who reviews them and replies.Gap found: the agent's work ends and nobody picks it upFixed: the agent hands refunds to a person?CustomerAI agentSupport agentRefund reviewNobody picks upOwner: a personCustomer replyBefore: the AI agent's refund step ends with nobody picking it up. After: the agent hands refunds to a person who reviews them and replies.Gap found: the agent's workends and nobody picks it upFixed: the agent handsrefunds to a person?CustomerAI agentSupport agentRefund reviewNobody picks upOwner: a personCustomer reply
Before: the AI agent's refund step ends with nobody picking it up. After: the agent hands refunds to a person who reviews them and replies.

At a glance

In this story
WhoThe two founders of an online homeware shop with a part-time team
Starting pointThe shop's website and its support process
UsedTeam types, step lanes, sub-processes, prompt for a developer or AI agent
What changedThe refund process has a clear point where the agent hands over to a person, and the developer had a precise brief

The problem

Most support questions were about delivery dates, and the AI agent handled them well. Refunds were different. Customers who wanted one were told to wait, and then nothing happened, because nobody had decided where the agent's job ended.

What they did

The founders mapped the shop's website and marked Customer Support as a hybrid team: part people, part AI agent. With a credit pack they opened the refund process and placed each step in a lane: the customer, the team, the AI agent and the courier as an outside party.

Laid out that way, the gap was plain. The agent's lane ended at 'refund requested' and no step in the team's lane picked it up. They broke the step into a sub-process and added a person who reviews refunds and replies to the customer.

They then exported the prompt in Requ for their developer and the AI coding agent the developer works with, so the change to the support agent matched the map.

Export: broadsheet PNG, handbook, fix list CSV, before and after, and the developer prompt
Export: broadsheet PNG, handbook, fix list CSV, before and after, and the developer prompt. Screenshot from a public example map, not the company in this story.
Take it to work · 25 s
Read the captions
  1. Take the map to work from Export.
  2. An operating handbook: who does what, for every member of staff.
  3. A fix list for Excel, Sheets, Jira or Linear.
  4. With credits: a prompt that hands the model to developers and AI coding agents.

What changed

Refund requests now reach a named person, and the agent tells customers who will reply. The map shows exactly where people and the agent hand work to each other.

Try it yourself

For teams working with AI agentsThe use case this story belongs to.

Questions

How do I map where an AI agent hands work to a person?
Mark the team as hybrid, then put each process step in a lane: the customer, the team, an AI agent or an outside party. Steps where the agent's lane ends with no next step are the gaps.
Can Requstory brief a developer or AI coding agent?
Yes, with credits. Export writes a prompt in Requ, the open process language, describing the teams, steps and handoffs to build against.
How deep can a process go?
Any step can become a sub-process, up to five levels deep.